Automated ejection fraction and risk stratification in cardiomyopathy patients with diverse LV geometry using 2D echocardiography.

Cardiomyopathy often alters left ventricular geometry (LVG), impairing cardiac function. We developed a deep learning (DL) model to estimate left ventricular ejection fraction (LVEF) from echocardiographic images while accounting for LVG variability and assessed prognostic factors across LVG subtypes. For all patients with cardiomyopathy, we computed LV volume on apical two- and four-chamber views processed with novel DeepLabV3+ algorithm and calculate EF using Simpson's method. The model was pre-trained on public data, then validated in 120 patients classified into concentric hypertrophy (CH), eccentric hypertrophy (EH), concentric remodeling (CR), or normal geometry (NG). Outcomes included cardiac death and heart failure rehospitalization, analyzed via logistic and LASSO regression within each LVG subtype. The model achieved high LV segmentation accuracy, with an overall Dice similarity coefficient of 90.07% and IoU of 82.17%. Subgroup analysis on A4C images showed Dice/IoU values of 92.49%/86.34% (NG), 88.91%/80.11% (CR), 88.81%/80.23% (CH), and 89.75%/81.59% (EH). The mean absolute error in LVEF estimation was 4.70%, and Bland-Altman analysis showed a mean bias of 0.95 ± 4.53% (95% limits, - 7.92% to 9.82%; P = 0.002) between AI-predicted and manual LVEF measurements. Subgroup analysis revealed r values of 0.794 (CR), 0.526 (CH), and 0.968 (EH). During follow-up, 20 patients experienced adverse outcomes. LASSO regression identified predicted LVEF, E/e' ratio, and age as significant predictors, with AUC values of 0.833 (CR), 0.695 (CH), and 0.938 (EH) for adverse outcomes prediction. This DL model provides accurate LVEF estimates across diverse LVG subtypes, offering a geometry-specific tool for clinical assessment and risk stratification in cardiomyopathy.